KR20220095092A - Forest fire susceptibility mapping method and apparatus using artificial intelligence - Google Patents

Abstract

Disclosed are a forest fire susceptibility map generation method using artificial intelligence and an apparatus using the same. The forest fire susceptibility map generation method using artificial intelligence comprises the steps of: (a) building a spatial database, wherein the spatial database includes forest fire data and forest fire-related factors; (b) analyzing a spatial correlation by deriving a counter variogram, a Moran index value, and a frequency ratio (FR) for the forest fire-related factor based on the spatial database; (c) training a forest fire susceptibility inference model by differently reflecting weights for the forest fire-related factor using the spatial database based on the results of the spatial correlation analysis; and (d) generating a forest fire susceptibility map for a target region by applying data on the target region to the trained forest fire susceptibility inference model. The present invention can increase network learning and inference accuracy of the inference model.

Description

Forest fire susceptibility mapping method and apparatus using artificial intelligence

The present invention relates to a method and apparatus for generating a map of forest fire vulnerability using artificial intelligence.

Forests are one of the important natural resources with positive status, economic and social effects. In recent years, population growth and industrial development have increased the use of natural resources, especially forests, to provide food and materials, and various human and natural factors have increased the risk of forest fires.

16 million hectares of the world's forest area are destroyed each year to increase agricultural land use/coverage, which is a global problem. After agriculture and urban activities, forest fires are the leading cause of ecosystem destruction.

Forest fires are caused by various factors such as camping in the forest, intentional forest fires to convert forests to agricultural land, lighting of fires, and drought. Therefore, there is a need for forest fire risk mapping and research to identify vulnerable areas.

The present invention is to provide a method and apparatus for generating a map of forest fire vulnerability using artificial intelligence.

Another object of the present invention is to provide a method and an apparatus for generating a forest fire vulnerability map using artificial intelligence that can be provided to a user in conjunction with geographic information by generating a forest fire vulnerability map through a plurality of inference models.

In addition, the present invention provides a method and apparatus for generating a forest fire vulnerability map using artificial intelligence that can increase network learning and inference accuracy of an inference model by reflecting the result of spatial correlation analysis using forest fire-related factors as a weight when learning an inference model is to provide

According to one aspect of the present invention, there is provided a method for generating a map of forest fire vulnerability using artificial intelligence.

According to an embodiment of the present invention, (a) constructing a spatial database, the spatial database includes forest fire data and forest fire-related factors; (b) analyzing the spatial correlation by deriving a half-variability value, a Moran's I index value, and a frequency ratio (FR) for the forest fire-related factors, respectively, based on the spatial database; (c) learning a forest fire vulnerability inference model by using the spatial database based on the spatial correlation analysis result to reflect different weights for the forest fire-related factors; and (d) generating a forest fire vulnerability map for the target area by applying data on the target area to the learned forest fire vulnerability inference model may be provided.

The inference model is an adaptive neuron-fuzzy (ANFIS) inference model, which is trained using a genetic algorithm (GA) and a simulated annealing algorithm (SA), and some training data in the spatial database The adaptive neuro-fuzzy inference model is initially trained with the genetic algorithm based on .

The forest fire-related factors include the year of the fire, fire radiative power (FRP), altitude, inclination angle and slope, distance to road, land use, distance to residence, rainfall, temperature, wind speed, wind direction and At least a plurality of soil types.

The inference model is a radial basis function-based ensemble model, but may be learned using an empirical competitive algorithm (ICA).

The wildfire-related factors are divided into a continuous data category and a discrete data category, the continuous data category and the discrete data category are assigned weights in different ways, and the discrete data category is weighted using the frequency ratio, , after the continuous data category is assigned to the middle of the corresponding category, weights for all values of forest fire-related factors corresponding to the continuous data category may be calculated using the radial basis function-based ensemble model.

The forest fire vulnerability map may be output through a user terminal in conjunction with geographic information (GIS).

According to another aspect of the present invention, an apparatus for generating a map of forest fire vulnerability using artificial intelligence is provided.

According to an embodiment of the present invention, a data configuration unit for constructing a spatial database - the spatial database includes forest fire data and forest fire-related factors; a spatial correlation analysis unit for analyzing spatial correlations by deriving a half-variability value, a Moran's I index value, and a frequency ratio (FR) for the forest fire-related factors, respectively, based on the spatial database; a learning unit configured to learn a forest fire vulnerability inference model by using the spatial database based on the spatial correlation analysis result to reflect different weights for the forest fire-related factors; and a vulnerability map generator configured to generate a forest fire vulnerability map for the target area by applying data on the target area to the learned forest fire vulnerability inference model.

By providing a method and apparatus for generating a forest fire vulnerability map using artificial intelligence according to an embodiment of the present invention, a forest fire vulnerability map can be generated through a plurality of inference models and provided to a user in conjunction with geographic information.

In addition, the present invention has an advantage in that it is possible to increase the network learning and inference accuracy of the inference model by reflecting the result of spatial correlation analysis using the forest fire-related factors as a weight when learning the inference model.

1 is a flowchart illustrating a method for generating a forest fire vulnerability map according to an embodiment of the present invention.
2 is a diagram illustrating a location of a wildfire in a target area according to an embodiment of the present invention.
3 is a diagram illustrating a forest fire-related factor according to an embodiment of the present invention.
Figure 4 is a diagram illustrating a result of spatial correlation analysis of the year of wildfire and wildfire radiated heat energy according to an embodiment of the present invention.
5 is a diagram illustrating a result of spatial correlation analysis of forest fire-related factors according to an embodiment of the present invention.
Figure 6 is a view showing a Moran's I index analysis results according to an embodiment of the present invention.
7 is a view showing a forest fire hot spot analysis result according to an embodiment of the present invention.
8 is a view showing a frequency ratio (FR) analysis results of forest fire-related factors according to an embodiment of the present invention.
9 is a diagram illustrating an objective function convergence result of a combining algorithm combining a GA algorithm and an SA algorithm according to an embodiment of the present invention.
10 is a diagram illustrating performance results for an adaptive neuron-fuzzy (ANFIS) inference model according to an embodiment of the present invention.
11 is a diagram illustrating a forest fire vulnerability map generated through an adaptive neuron-fuzzy (ANFIS) inference model according to an embodiment of the present invention.
12 is a diagram illustrating a fitness result of an ensemble model for forest fire-related factors according to an embodiment of the present invention.
13 is a diagram illustrating a forest fire vulnerability map generated through an ensemble model according to an embodiment of the present invention.
14 is a diagram illustrating a user terminal display screen displayed by mapping a forest fire vulnerability map for a target location to geographic information (GIS) according to an embodiment of the present invention.
15 is a diagram illustrating an ROC curve result according to an embodiment of the present invention.
16 is a block diagram schematically illustrating an internal configuration of an apparatus for generating a map of forest fire vulnerability according to an embodiment of the present invention.

As used herein, the singular expression includes the plural expression unless the context clearly dictates otherwise. In this specification, terms such as "consisting of" or "comprising" should not be construed as necessarily including all of the various components or various steps described in the specification, some of which components or some steps are It should be construed that it may not include, or may further include additional components or steps. In addition, terms such as "...unit" and "module" described in the specification mean a unit that processes at least one function or operation, which may be implemented as hardware or software, or a combination of hardware and software. .

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

1 is a flowchart illustrating a method for generating a forest fire vulnerability map according to an embodiment of the present invention, FIG. 2 is a diagram illustrating the location of a wildfire occurrence in a target area according to an embodiment of the present invention, and FIG. 3 is a diagram of the present invention It is a diagram illustrating a forest fire-related factor according to an embodiment, and FIG. 4 is a diagram illustrating a result of spatial correlation analysis of the wildfire occurrence year and wildfire radiant heat energy according to an embodiment of the present invention, and FIG. 5 is a diagram of the present invention It is a diagram illustrating a result of spatial correlation analysis of forest fire-related factors according to an embodiment, FIG. 6 is a diagram illustrating a Moran's I index analysis result according to an embodiment of the present invention, and FIG. 7 is an embodiment of the present invention It is a view showing the analysis result of a forest fire hotspot according to an example, and FIG. 8 is a diagram showing the analysis result of a frequency ratio (FR) of a forest fire-related factor according to an embodiment of the present invention, and FIG. 9 is an embodiment of the present invention It is a view showing the result of objective function convergence of the combining algorithm combining the GA algorithm and the SA algorithm according to , FIG. 11 is a diagram illustrating a forest fire vulnerability map generated through an adaptive neuron-fuzzy (ANFIS) inference model according to an embodiment of the present invention, and FIG. 12 is a forest fire of an ensemble model according to an embodiment of the present invention. It is a diagram illustrating a fitness result for a related factor, and FIG. 13 is a diagram illustrating a forest fire vulnerability map generated through an ensemble model according to an embodiment of the present invention, and FIG. 14 is a target according to an embodiment of the present invention. It is a diagram illustrating a user terminal display screen displayed by mapping a map of forest fire vulnerability for a location to geographic information (GIS), and FIG. 15 is a diagram illustrating an ROC curve result according to an embodiment of the present invention. In step 110, the apparatus 100 for generating a map of forest fire vulnerability configures a spatial database.

Here, the spatial database may include forest fire data and forest fire-related factors.

This will be described in more detail.

The forest fire vulnerability map generating apparatus 100 may select a forest fire occurrence location using a satellite image, and may acquire survey data for the forest fire occurrence location.

For example, in one embodiment of the present invention, forest fire data using satellite imagery and survey data from the Chaharmahal and Bakhtiari Departments of Natural Resources, MODIS hotspot products (http://earthdata.nasa.gov/firms) from 2013 to 2018 in one embodiment of the present invention can be collected. That is, about 262 wildfire data were collected in the target area with a 1km buffer.

The forest fire location-related data collected in this way was randomly classified for training and validation of the inference model at a ratio of 7:3.

Forest fire-related factors can be constructed by considering topographical factors, socioeconomic factors, and climatic factors that influence wildfires. Hereinafter, this will be briefly described.

1. Geographical factors

Topographical variables that affect the occurrence of wildfires can be elevation, inclination angle, and slope. Topographical features and wind direction can also affect the likelihood of a wildfire in an area.

In an embodiment of the present invention, a digital elevation model (DEM) was prepared with a spatial resolution of 30 m from the STER satellite image, and maps of elevation, inclination angle and slope were extracted through ArcGIS10.3 software.

Altitude change affects temperature and vegetation, and it can be seen that the possibility of wildfire occurrence is inversely proportional to the high altitude because the high-altitude area has low temperature and high humidity (Fig. 3 (a))

Inclination angle also affects the direction and extent of wildfires. In general, the degree of damage is higher on steep slopes (Fig. 3(b)).

The slope is affected by the amount of sunlight and heat associated with drying the vegetation in the area. The south and east directions have a relatively high rate of sunlight exposure, which makes them more vulnerable to wildfire (Fig. 3 (c)).

2. Socioeconomic factors

As the population has increased in recent decades, forest areas are being damaged for supply, fuel and agriculture, and forest fires caused by some tourists are also on the rise.

In one embodiment of the present invention, three factors, namely the distance to the road, land use, and the distance to the residence, were set as socioeconomic factors.

In general, wildfires are more likely to occur near roads. Therefore, a road map of the target area was created using a 1:100,000 topographic map, and buffers for roads were defined at specified intervals (see Fig. 3(f)))

Another influencing factor affecting wildfires is proximity to settlements (rural and nomadic settlements). Because farmers depend on forests and land to meet their needs, they can cause both deliberate and unintentional wildfires.

The distance from the settlement map was created using a 1:100,000 topographic map (Fig. 3(g)). Landsat-7 images from 2013 and 2018 were used to map land use and residence.

In an embodiment of the present invention, a land use and residential map was created using 400 target points (training points) collected by GPS (Global Positioning System), 70% of which was used for training, and the remaining 30% was used for accuracy. was used to evaluate

A land use/cover map was obtained with 91% accuracy using the maximum likelihood algorithm and divided into five classifications such as forest, agriculture, pasture, wilderness, and urban area (see Fig. 3(d)).

3. Climate Factors

Climatic conditions directly or indirectly affect wildfires and indirectly affect the type and density of vegetation. In an embodiment of the present invention, annual average data on rainfall, temperature, wind speed and wind direction were obtained from 14 meteorological stations in Chaharmahal and Bahtiari provinces from 2013 to 2018.

The obtained climatic factor-related data were prepared as rainfall and temperature maps using the Kriging interpolation method.

In addition, wind effect maps were constructed using Windward and Leeward indices of the Automated Geoscientific Analysis (SAGA)-GIS software system. Inputs for these indicators were DEM, speed and wind direction. In general, areas with high temperatures and low precipitation affect the occurrence of forest fires (refer to (h) ~ (j) of FIG. 3).

In addition, the year of the wildfire, the fire radiative power (FRP: fire radiative power), the soil type, and the like may be constructed as factors related to the forest fire.

In operation 115 , the apparatus 100 for generating a forest fire vulnerability map calculates a forest fire frequency and a forest fire period.

Wildfire frequency and fire frequency are used to analyze the location of wildfires. Wildfire frequency is defined as the number of wildfires occurring within a target area and period.

A forest fire cycle is defined as the time required to burn an area equal to the area of the target area.

In the event of a wildfire, the entire area may not burn, and some places may burn several times or not at all.

In an embodiment of the present invention, the forest fire cycle may be calculated using Equation 1.

는 전체 영역을 나타내고,

는 년 수를 나타내며,

는 전체 영역 화재 사이클(fire cycle=burned)을 나타낸다. here,

represents the entire area,

represents the number of years,

is the total area fire cycle (fire cycle=burned).

In step 120, the apparatus 100 for generating a forest fire vulnerability map calculates the half-variability value, the Moran's I index, and the frequency ratio (FR) for the wildfire-related factors to analyze the spatial correlation.

Each of these will be described.

As already mentioned above, forest fire-related factors include the year of the fire, fire radiative power (FRP), altitude, inclination angle and slope, distance to road, land use, distance to dwelling, rainfall, temperature, wind. It can be speed, wind direction and soil type.

The apparatus 100 for generating a forest fire vulnerability map calculates half-variance values for some of the forest fire-related factors, respectively.

For the convenience of understanding and explanation, the degree of half-variability will be briefly described.

Half-variability is a measure of the correlation between data at a constant distance. Therefore, when the distance is close, the values are similar, so it has a small value, and it is general that the value appears larger as the distance increases.

,

)에서 확률 과정(Z)에 의해 가정된 값들 사이에서의 연관성은 변이도 함수로 표현되며, 이에 대한 추정량은 수학식 2와 같이 계산될 수 있다. For example, two points within the region of interest (

,

), the association between the values assumed by the probabilistic process Z is expressed as a variability function, and an estimator for this can be calculated as in Equation (2).

는 실험적 반변이도의 값을 나타내며, 평균이 상수라는 가정에서 적률법을 이용한 불편추정량이다. here,

represents the value of the experimental half-variability, and is an unbiased estimator using the moment method under the assumption that the mean is constant.

와

는 두 지점(i)와 (i+h)의 로컬 변수를 나타낸다. In addition,

Wow

denotes the local variables of two points (i) and (i+h).

Half-variability is modeled by range, still, and nugget.

range is the distance at which the variogram reaches a fixed point and approaches the horizon, and specifies the range over which data can be used to estimate the value of an unknown variable, beyond which there is no longer spatial continuity. , the corresponding sample works independently.

The sill parameter is a constant value reached in the variability range parameter. This is equal to the variance of the total number of samples used to calculate the variability.

A nugget representing a variable structure is called a nugget, except that at the origin of the coordinates (h = 0), the quantity of the variogram should be random and ideally zero.

When half-variability values are derived for some of the forest fire-related factors, spatial dependence (SD) may be calculated using parameters used to derive half-variability values.

For example, the spatial dependence (SD) can be calculated using the nugget and partial still parameters, as shown in Equation (3).

If SD is less than 25%, the variable has strong spatial dependence, between 25% and 75%, the variable has moderate spatial dependence, and if the SD is greater than 75%, the variable has weak spatial dependence.

The forest fire vulnerability map generating apparatus 100 calculates the half-variability value using the year of the wildfire and fire radiative power (FRP) as parameters, as shown in Tables 1 and 4 .

[Table 1]

The value of the nugget parameter of the fire year was smaller than that of FRP, and the two parameters of rnage and partial still of FRP were higher.

According to the SD index results, it can be seen that the year of fire is more spatially dependent than FRP. Therefore, it can be seen that the use of wildfire year data in the preparation of forest fire vulnerability maps can achieve better results than FRP.

Table 2 and Figure 5 show the semivariance values and spatial dependence (SD) of forest fire-related factors such as altitude, slope angle and slope, distance to road, land use, distance to dwelling, rainfall, temperature, wind speed, and wind direction. The result of calculating , is shown.

[Table 2]

Referring to Table 2 and FIG. 5 , the values of the nugget parameters were the lowest for soil, rainfall, and distance to a settlement, and the values of the nugget parameters were high for slopes, inclination angles, and wind influence variables.

Also, in the case of range parameters, it was found that temperature, land use/cover, and distance to road variables had high values, and soil and slope inclination parameters had low values.

For the sill parameters, distance to road, temperature, and land use/cover variables were high, and altitude, soil and distance to settlement variables were low.

In addition, SD index results showed that distance to settlement, soil and precipitation variables were high, and wind influence, inclination angle and slope variables were low.

As described above, the apparatus 100 for generating a forest fire vulnerability map may calculate a semivariance value and spatial dependence (SD) of the wildfire-related factors.

Also, the apparatus 100 for generating a forest fire vulnerability map calculates Moran's I indices for some of the forest fire-related factors, respectively.

Moran's I indices include general and regional Moran's I indicators, which are considered essential tools for examining spatial relationships between data.

The general Moran's I index indicators range in value from -1 to +1. A quantity of +1 indicates the highest positive spatial correlation, a quantity of -1 indicates the highest negative spatial correlation, and 0 indicates the maximum degree of random distribution of values.

The generic Moran's I index indicator shows whether there is a general spatial correlation in the data. A regional Moran's I index indicator can indicate the degree of spatial correlation of a variable at any point. General Moran's I and local Moran's I indices can be calculated using Equations 4 and 5.

는 I 위치의 변수값을 나타내고,

는 j위치의 변수값을 나타내며,

는 기술된 값의 평균을 나타내며,

는 위치 I의 값을 나타내고,

는 i와 j 사이의 공간적 공동 가중치(co-weight)를 나타내며, n은 분규(complication)의 총 수를 나타낸다. In Equations 4 and 5, Z represents the average value of the variable z,

represents the variable value at the I position,

represents the variable value at the j position,

represents the average of the stated values,

represents the value of position I,

denotes the spatial co-weight between i and j, and n denotes the total number of complications.

The calculation results of Moran's I index are shown in Table 3 and FIG. 6 .

Moran's I index values for the year of fire and FRP parameters were 0.830061 and 0.904367, respectively, and it can be seen that the year of fire and FRP variables show a cluster distribution.

The results of clustering and analyzing forest fire hotspot areas using the calculated Moran's I index are as shown in FIG. 7 . As a result of the analysis of the Moran's I index for the year of the wildfire and the FRP factor, it was analyzed that the wildfire hotspot was in the northwest part, and the rest of the area was a cold spot and a temporary spot.

In addition, the apparatus 100 for generating a forest fire vulnerability map derives a frequency ratio (FR) for some of the forest fire-related factors.

For the convenience of understanding and explanation, the FR model will be briefly described.

The relationship between wildfires and wildfire-related factors is inferred from the areas where wildfires occurred or did not occur and from the wildfire-related factors. To express this quantitatively, the FR model is used.

FR is the ratio of the forest fire occurrence area by class of each forest fire-related factor divided by the ratio of each forest fire-related factor class to the total area.

For example, FR can be calculated using Equation (6).

는 해당 인자(x)의 i번째 클래스의 산불이 발생한 픽셀의 개수를 나타내고,

는

의 전체 픽셀 개수를 나타내며, m과 n은 카테고리의 개수와 기준의 전체 수를 나타낸다. here,

represents the number of pixels in the i-th class of the corresponding factor (x) where wildfires occurred,

Is

represents the total number of pixels, and m and n represent the number of categories and the total number of criteria.

In other words, FR is the ratio of each factor's proportion of forest fire occurrence area. If the value of FR is greater than 1, it indicates a high correlation between the forest fire and the factors affecting it, and if it is lower than 1, it indicates a low correlation.

After dividing the 10 forest fire influencing factors for each class, the FR value was calculated, and the calculation result is shown in FIG. 8 .

As shown in FIG. 8 , the results for the spatial relationship between forest fire influencing factors and forest fire locations are shown.

As a result of the altitude, the highest FR value was 2.81, indicating that it is related to the altitude class above 2600m. In addition, in the case of the slope, the maximum FR value was 1.67, which was found to be related to the 15 to 20 degree class. In the case of the slope, the highest FR value was 1.23, indicating that the class was in the southwest direction.

For the distance factor to the road, the 300 ~ 600m class had the highest FR value of 1.42, and for the distance factor to the settlement (residential), the 0 ~ 500m class showed the highest FR value with 2.096, and in the soil type factor, Entiso group class had the highest FR value with 1.39.

Also, in the case of rainfall, the 750mm class had the highest FR value of 5.96, and in the case of temperature, the 8 degree class had the highest FR value of 4.06, and in the case of the wind influence factor, the FR value was the highest at 2.21 in the class larger than 1.2. , and for land use factors, the forest class had the highest FR value of 1.8.

In step 125, the apparatus 100 for generating a forest fire vulnerability map learns a forest fire vulnerability inference model based on the spatial database by reflecting different weights for forest fire-related factors based on the spatial correlation analysis result.

According to an embodiment of the present invention, the forest fire vulnerability inference model may be at least one of an adaptive neuron-fuzzy (ANFIS) inference model and a radial basis function-based ensemble model.

ANFIS was utilized with descending gradient and least squares techniques. Least squares method is effective in optimizing parameters, and the antecedent parameters are very important in the modeling of ANFIS inference models. The conventional method is not suitable due to the high complexity of the gradient calculation and the nonlinear presence of the antecedent parameters in the output. Therefore, in one embodiment of the present invention, in training the adaptive neuron-fuzzy (ANFIS) inference model, the accuracy is improved by combining a genetic algorithm (GA) and a simulated annealing algorithm (SA). raised

The GA algorithm is based on the theory of evolution and is one of the first and most popular metaheuristic algorithms. The basis of the GA algorithm is to transform each set of answers into a binary encoder called a chromosome. Each chromosome in the GA algorithm represents a search space location and a possible solution to the problem. The GA algorithm begins with the generation of an original set of randomized alternatives called a population. Iterations of the GA algorithm produce new sets of chromosomes called generations, the amount of which is determined by a function of fitness over each generation. Reproductive processes include fusion and mutation in which genetic operators are applied to chromosomes. The fitness function evaluates each chromosome, the genetic operator selects it, and selects the optimal chromosome to pass the optimal solution to the problem to the next generation.

The SA algorithm starts with the original solution and moves to the adjacent solution in an iterative loop. If the neighbor's answer is better than the original answer, the SA algorithm sets it as the original answer, and if not, the original algorithm accepts the original answer along with the likelihood. This is repeated several times at each temperature as the temperature gradually decreases. The initial stage sets the temperature to accommodate the worse response, and as the temperature is gradually decreased, the probability of accepting the worse solution in the final stage decreases so that the algorithm converges to a good solution.

In an embodiment of the present invention, a forest fire vulnerability inference model can be learned using a combination algorithm that combines the GA algorithm and the SA algorithm.

According to an embodiment of the present invention, the initial model of the forest fire vulnerability inference model may be configured and trained with a GA algorithm. When the termination condition of the GA algorithm is reached, the forest fire vulnerability inference model can be learned by applying the SA algorithm.

Here, the termination condition of the GA algorithm is that the best chromosome found for 5 generations does not change, or if the GA algorithm, which is half the termination condition of the GA algorithm, is repeated 10 times, the SA algorithm is activated and the forest fire vulnerability inference model can be trained.

To this end, the best solution for the GA algorithm is selected as the first solution for the SA algorithm, and the remaining calculations can be performed.

The operations of the GA algorithm and the SA algorithm are as follows.

Step 1 - The original chromosome population is selected.

Stage 2 - The current generation is incremented using mutation and union operators.

Step 3—If the best chromosome is not found in 5 consecutive iterations, or if the iteration has been repeated more than 10 times, which is half of the termination condition, the best current chromosome is considered as input to the SA algorithm.

Step 4 - Train the ANFIS using the SA algorithm.

The objective function for training the ANFIS in the combination algorithm combining the GA algorithm and the SA algorithm is as shown in Equations 7 and 8.

Here, t represents target data and y represents input data.

The parameters used in the GA algorithm and the SA algorithm of the adaptive neuron-fuzzy (ANFIS) inference model are shown in Table 5, and the convergence result of the objective function by the combination algorithm is shown in FIG. 9 .

[Table 5]

Based on the joint algorithm for the adaptive neuron-fuzzy (ANFIS) inference model, fire locations with a value of 1 and non-fire locations with a value of 0 were randomly assigned as training data.

The results of the RMSE, MSE, and MAE parameters of the joint algorithm on the training data were 0.2672, 0.0713, and 0.1735, respectively, and the results of the RMSE, MSE, and MAE parameters of the joint algorithm on the validation data were 0.4286, 0.1837, and 0.2963, respectively. .

The convergence result for the objective function of the combining algorithm is as shown in FIG. 9 , and FIG. 10 shows the performance of the combining algorithm on the training data and the test data.

A forest fire vulnerability map was generated after training a network of an adaptive neuron-fuzzy (ANFIS) inference model with a combined algorithm. It can be seen that the forest fire vulnerability map is divided into very low, low, medium, high and very high risk classifications (see FIG. 11 ).

As another example, the forest fire vulnerability inference model is a radial basis function-based ensemble model, and may be learned using an empirical competitive algorithm (ICA).

A brief description of the radial basis function-based ensemble model is as follows.

The radial basis function-based ensemble model is a spline interpolation method used when the modeling process takes a lot of time. Unlike neural network-based prediction methods, this model provides function estimation with much lower computation and higher accuracy. An ensemble model based on a radial basis function can estimate a new value at a new location in the sample by applying a reference point.

For example, the radial basis function-based ensemble model may calculate an interpolation function as shown in Equation 9 based on the sum of data.

는 두 지점의 유클리드 거리를 나타내고,

는 가중치를 나타내고,

는 방사 기저 함수를 나타낸다. here,

represents the Euclidean distance between two points,

represents the weight,

is the radiative basis function.

According to an embodiment of the present invention, the radial basis function-based ensemble model may be trained using an empirical competitive algorithm (ICA).

First, the present invention classifies forest fire-related factors into a continuous data category and a discrete data category. For example, a continuous data category may include factors such as slope angle, elevation, slope, temperature, wind influence, rainfall, distance to road, and distance to a settlement, while a discrete data category may include soil and land use factors.

In this way, after dividing forest fire-related factors into continuous data categories and discrete data categories, in the case of discrete data categories, learning is performed using the weights of the FR model, and the continuous data categories are radial basis functions after assigning the median of the factors as weights. Calculate the weights for all values using the underlying ensemble model.

According to an embodiment of the present invention, the interpolation method of the ensemble model based on the radial basis function may be learned to optimize the objective function according to the training data at low tide by using the center of each class of each forest fire-related factor as training data. As described above, the radial basis function-based ensemble model may be trained using an empirical competitive algorithm (ICA), and the objective function may be calculated as in Equation (10).

및

는 커널 함수의 중간과 가중치의 편차 및 계수를 나타낸다. where m,

and

denotes the deviation and coefficient of the mean and weight of the kernel function.

12 shows the best fit for each wildfire-related factor using the ensemble model. 12 , the horizontal axis is the value of each forest fire-related factor, and the vertical axis is the FR value.

Table 6 shows the parameters used in the empirical competitive algorithm (ICA) used to train the radial basis function-based ensemble model.

The forest fire vulnerability map generated by the learned radial basis function-based ensemble model is shown in FIG. 13 .

In step 130 , the apparatus 100 for generating a forest fire vulnerability map generates a forest fire vulnerability map for the target area by applying data on the target area to the learned forest fire vulnerability inference model. At this time, as shown in FIG. 13 , the forest fire vulnerability map may be provided to the user terminal in conjunction with geographic information.

The wildfire vulnerability map output to the user terminal may be displayed by mapping the wildfire vulnerability map for the user location or the target location selected by the user to the geographic information (GIS) in conjunction with the geographic information (GIS) (see FIG. 14).

In addition, as described above, the apparatus 100 for generating a forest fire vulnerability map according to an embodiment of the present invention generates a forest fire vulnerability map through a forest fire vulnerability inference model. A forest fire vulnerability map can be generated for each of the adaptive neuron-fuzzy (ANFIS) inference model trained by the adaptive neuron-fuzzy (ANFIS) inference model and the radial basis function-based ensemble model trained by the empirical competition algorithm.

Therefore, the forest fire vulnerability map generated by each of the learned adaptive neuron-fuzzy (ANFIS) inference model and the trained radial basis function-based ensemble model is either displayed or any one is selectively displayed by the user interface through the user terminal. may also provide

In one embodiment of the present invention, 79 wildfire occurrence locations (value of 1) and 79 random points (value of 0) were used to evaluate the wildfire fragility map (FFSM) using the ROC curve. 14 and Table 7 show the results for ROC-AUC. According to the results, the accuracies of the trained adaptive neuron-fuzzy (ANFIS) inference model applied with the combining algorithm and the radial basis function-based ensemble model with the empirical competition algorithm are 0.903 and 0.878, respectively.

[Table 7]

Table 8 shows the results for the RMSE and MAE parameters of the learned adaptive neuron-fuzzy (ANFIS) inference model to which the joint algorithm is applied. As a result, the RMSE and MAE parameter values for the training data and the test data were 0.267, 0.173, 0.428, and 0.296, respectively.

[Table 8]

16 is a block diagram schematically illustrating an internal configuration of an apparatus for generating a map of forest fire vulnerability according to an embodiment of the present invention.

Referring to FIG. 16 , the apparatus 100 for generating a forest fire vulnerability map according to an embodiment of the present invention includes a communication unit 1610 , a data configuration unit 1615 , a spatial context analysis unit 1620 , a learning unit 1625 , It is configured to include a forest fire vulnerability map generator 1630 , a memory 1635 , and a processor 1640 .

The communication unit 1610 is a means for transmitting and receiving data with another device (eg, a user terminal, etc.) through a communication network.

The data configuration unit 1615 is a means for building a spatial database. Since this is the same as that described above, the overlapping description will be omitted.

The spatial context analysis unit 1620 derives the half-variability value, the Moran's I index value, and the frequency ratio (FR) for the wildfire-related factors based on the spatial database, respectively, and uses them to analyze the spatial correlation. .

Since this is the same as that described above, the overlapping description will be omitted.

The learning unit 1625 is a means for learning the forest fire vulnerability inference model by differently reflecting the weights for the forest fire-related factors using a spatial database based on the spatial correlation analysis result.

As described above, the forest fire vulnerability inference model is an adaptive neuro-fuzzy (ANFIS) inference model trained by a combined algorithm that combines the GA algorithm and the SA algorithm, and a radial basis function-based ensemble model trained using an empirical competition algorithm. can

The forest fire vulnerability inference model may generate a forest fire vulnerability map for a target area using each of the adaptive neuro-fuzzy inference model and the radial basis function-based ensemble model, respectively.

Since this is the same as that described above, the overlapping description will be omitted.

The memory 135 is a means for storing instructions for performing the method for generating a forest fire vulnerability map using artificial intelligence according to an embodiment of the present invention.

The processor 1640 includes internal components (eg, a communication unit 1610 , a data structure unit 1615 , and a spatial context relationship analysis unit 1620 ) of the apparatus 100 for generating a forest fire vulnerability map according to an embodiment of the present invention. ), the learning unit 1625, the forest fire vulnerability map generation unit 1630, the memory 1635, etc.).

The apparatus and method according to an embodiment of the present invention may be implemented in the form of program instructions that can be executed through various computer means and recorded in a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, and the like, alone or in combination. The program instructions recorded on the computer readable medium may be specially designed and configured for the present invention, or may be known and available to those skilled in the art of computer software. Examples of the computer-readable recording medium include magnetic media such as hard disks, floppy disks and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic such as floppy disks. - Includes magneto-optical media and hardware devices specially configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like. Examples of program instructions include not only machine language codes such as those generated by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like.

The hardware devices described above may be configured to operate as one or more software modules to perform the operations of the present invention, and vice versa.

So far, the present invention has been looked at focusing on the embodiments thereof. Those of ordinary skill in the art to which the present invention pertains will understand that the present invention can be implemented in a modified form without departing from the essential characteristics of the present invention. Therefore, the disclosed embodiments should be considered in an illustrative rather than a restrictive sense. The scope of the present invention is indicated in the claims rather than the foregoing description, and all differences within the scope equivalent thereto should be construed as being included in the present invention.

100: Forest Fire Vulnerability Map Generator
1610: communication department
1615: data structure part
1620: spatial correlation analysis unit
1625: Study Department
1630: Forest Fire Vulnerability Map Generation Unit
1635: memory
1640: Processor

Claims (12)

(a) constructing a spatial database, the spatial database including wildfire data and wildfire-related factors;
(b) analyzing the spatial correlation by deriving a half-variability value, a Moran's I index value, and a frequency ratio (FR) for the forest fire-related factors, respectively, based on the spatial database;
(c) learning a forest fire vulnerability inference model by using the spatial database based on the spatial correlation analysis result to reflect different weights for the forest fire-related factors; and
(d) generating a forest fire vulnerability map for the target area by applying data on the target area to the learned forest fire vulnerability inference model;
According to claim 1,
The inference model is an adaptive neuron-fuzzy (ANFIS) inference model, which is learned using a genetic algorithm (GA) and a simulated annealing algorithm (SA),
The adaptive neuro-fuzzy inference model is initially trained with the genetic algorithm based on some training data in the spatial database, and when the termination condition of the genetic algorithm is reached, the simulated annealing algorithm is activated and the adaptive neuro-fuzzy reasoning A method of generating a forest fire vulnerability map, characterized in that the model is additionally trained.
According to claim 1,
The forest fire-related factors include: the year of the fire, fire radiative power (FRP), altitude, inclination angle and slope, distance to road, land use, distance to residence, rainfall, temperature, wind speed, wind direction and A method for generating a forest fire vulnerability map, characterized in that at least a plurality of soil types are present.
According to claim 1,
The inference model is a radial basis function-based ensemble model, a method for generating a forest fire vulnerability map, characterized in that it is learned using an empirical competitive algorithm (ICA: imperialist competitive algorithm).
5. The method of claim 4,
The wildfire-related factors are divided into a continuous data category and a discrete data category,
wherein the continuous data category and the discrete data category are weighted in different ways,
The discrete data category is assigned a weight using the frequency ratio, and the continuous data category is assigned to the middle of the corresponding category, and then a wildfire related factor corresponding to the continuous data category using the radial basis function-based ensemble model. A method of generating a forest fire vulnerability map, characterized in that weights for all values of are calculated.
According to claim 1,
The forest fire vulnerability map generating method, characterized in that it is output through a user terminal in conjunction with geographic information (GIS).
A computer-readable recording medium recording a program code for performing the method according to any one of claims 1 to 6.
a data configuration unit for constructing a spatial database, wherein the spatial database includes forest fire data and forest fire-related factors;
a spatial correlation analysis unit for analyzing spatial correlations by deriving a half-variability value, a Moran's I index value, and a frequency ratio (FR) for the forest fire-related factors, respectively, based on the spatial database;
a learning unit configured to learn a forest fire vulnerability inference model by using the spatial database based on the spatial correlation analysis result to reflect different weights for the forest fire-related factors; and
and a vulnerability map generator configured to generate a forest fire vulnerability map for the target area by applying data on the target area to the learned forest fire vulnerability inference model.
9. The method of claim 8,
The inference model is an adaptive neuro-fuzzy (ANFIS) inference model, which is learned using a genetic algorithm (GA) and a simulated annealing algorithm (SA),
The adaptive neuro-fuzzy inference model is initially trained with the genetic algorithm based on some training data in the spatial database, and when the termination condition of the genetic algorithm is reached, the simulated annealing algorithm is activated and the adaptive neuro-fuzzy reasoning A device for generating a map of forest fire vulnerability, characterized in that the model is additionally trained.
9. The method of claim 8,
The inference model is a radial basis function-based ensemble model, but forest fire vulnerability map generating apparatus, characterized in that it is learned using an empirical competitive algorithm (ICA: imperialist competitive algorithm).
11. The method of claim 10,
The wildfire-related factors are divided into a continuous data category and a discrete data category,
wherein the continuous data category and the discrete data category are weighted in different ways,
The discrete data category is assigned a weight using the frequency ratio, and the continuous data category is assigned to the middle of the corresponding category, and then a wildfire related factor corresponding to the continuous data category using the radial basis function-based ensemble model. A device for generating a map of forest fire vulnerability, characterized in that weights for all values of are calculated.
9. The method of claim 8,
The forest fire vulnerability map generating apparatus, characterized in that the map is output through a user terminal in conjunction with geographic information (GIS).

Images